Go To MPAI-MMC AI Modules

1     Function 2     Reference Model 3     Input/Output Data
4     SubAIMs 5     JSON Metadata 6     Profiles
7     Reference Software 8     Conformance Texting 9     Performance Assessment

1     Functions

Automatic Speech Recognition (MMC-ASR):

Receives Language Selector Signalling the language of the speech.
Auxiliary Text Text that may be used to provide context information.
Speech Object Speech to be recognised.
Speaker ID ID of speaker uttering speech.
Speech Overlap Data type providing information of speech overlap.
Speaker Time Time during which the speech is to be recognised.
Produces Recognised Text (Also called text transcript).

Recognised Text can be a Text Segment or just a string.

2     Reference Model

Figure 1 depicts the Reference Model of the Automatic Speech Recognition (MMC-ASR) AIM.

Figure 1 – The Automatic Speech Recognition (MMC-ASR) AIM

3    Input/Output Data

Table 1 specifies the Input and Output Data of the Automatic Speech Recognition (MMC-ASR) AIM.

Table 1 – I/O Data of the Automatic Speech Recognition (MMC-ASR) AIM

Input Description
Language Selector Selects input language
Auxiliary Text Text Object with content related to Speech Object.
Speech Object Speech Object emitted by Entity
Speaker ID Identity of Speaker
Speech Overlap Times and IDs of overlapping speech segments
Speaker Time Time during which Speech is recognised
Output Description
Recognised Text Output of the Automatic Speech Recognition AIM, a Text Segment or just a string.

4     SubAIMs

No SubAIMs.

5     JSON Metadata

https://schemas.mpai.community/MMC/V2.3/AIMs/AutomaticSpeechRecognition.json

6     Profiles

No Profiles

7     Reference Software

7.1    Disclaimers

  1. This MMM-ASR Reference Software Implementation is released with the BSD-3-Clause licence.
  2. The purpose of this Reference Software is to demonstrate a working Implementation of MMC-ASR, not to provide a ready-to-use product.
  3. MPAI disclaims the suitability of the Software for any other purposes and does not guarantee that it is secure.
  4. Use of this Reference Software may require acceptance of licences from the respective repositories. Users shall verify that they have the right to use any third-party software required by this Reference Software.

7.2    Guide to the ASR code #1

The code takes Speech Objects from MMC-AUS and generates Text Segments (called text transcripts). It uses the whisper-large-v3 model to convert an input Speech Object (speaker’s turn) into a Text Segment (here called text transcript). Disfluencies (e.g., repetitions, repairs, filled pauses) are often omitted. The Whisper reference document is available.

The MMC-ASR Reference Software is found at the MPAI gitlab site. Use of this AI Modules is for developers who are familiar with Python, Docker, RabbitMQ, and downloading models from HuggingFace. The Reference Software contains:

  1. src: a folder with the Python code implementing the AIM
  2. Dockerfile: a Docker file containing only the libraries required to build the Docker image and run the container
  3. requirements.txt: dependencies installed in the Docker image
  4. README.md: commands for cloning https://huggingface.co/openai/whisper-large-v3

Library: https://github.com/linto-ai/whisper-timestamped

7.3    Guide to the ASR code #2

Use of this AI Modules is for developers who are familiar with Python and downloading models from HuggingFace,

A wrapper for the Whisper NN Module:

  1. Manages input files and parameters: Speech Object
  2. Performs Speech Recognition on each Speech Object by executing the Whisper Module.
  3. Outputs Recognised Text.

The MMC-ASR Reference Software is found at the NNW gitlab  site (registration required). It contains:

  1. The python code implementing the AIM.
  2. The required libraries are: pytorch and transformers (HuggingFace).

7.4    Acknowledgements

This version of the MMC-ASR Reference Software

  1. #1 has been developed by the MPAI AI Framework Development Committee (AIF-DC).
  2. #2 has been developed by the MPAI Neural Network Watermarking Development Committee (NNW-DC).

8     Conformance Testing

Table 2 provides the Conformance Testing Method for MMC-ASR AIM.

If a schema contains references to other schemas, conformance of data for the primary schema implies that any data referencing a secondary schema shall also validate against the relevant schema, if present and conform with the Qualifier, if present.

Table 2 – MMC-ASR AIM Conformance Testing

Input Language Selector Shall validate against the Language Selector part of the schema.
Auxiliary Text Shall validate against the Text Object schema.
Text Data shall conform with the Text Qualifier.
Speech Object Shall validate against the Speech Object schema.
Speech Data shall conform with the Speech Qualifier.
Speaker ID Shall validate against the Instance ID schema.
Speech Overlap Shall validate against the Speech Overlap schema.
Speaker Time Shall validate against the Time schema.
Output Text Object Shall validate against the Text Object schema.
Text Data shall conform with the Text Qualifier, e.g. output language shall be that indicated by the Language Selector,

Table 3 provides an example of MMC-ASR AIM Conformance Testing.

Table 3 – An example of MMC-ASR AIM Conformance Testing

Input Data Data Format Input Conformance Testing Data
Speech Object .wav All input Speech files to be drawn from Speech files.
Output Data Data Format Output Conformance Testing Criteria
Recognised Text Unicode All Text files produced shall conform with Text files.

9     Performance Assessment

Performance Assessment of an ASR Implementation (ASRI) can be performed for a language for which there is a dataset of speech segments of various durations with corresponding Transcription Text. An MMC-ASR AIM Performance Assessment Report shall be based on the following steps and specify the input dataset used.

For each Recognised Text produced by the ASRI being Assessed for Performance in response to a speech segment provided as input:

  1. Compare the Recognised Text with the Transcription Text
  2. Compute the Word Error Rate (WER) defined as the sum of deletion, insertion, and substitution errors in the Recognised Text compared to the Transcription Text, divided by the total number of words in the Transcription Text.

This code can be used to compute the WER.

Performance Assessment of an ASRI for a language in a Performance Assessment Report is defined as “The WER computed on all speech segments included in the reported dataset”.

 

Go To MPAI-MMC AI Modules